Lymphoma is a malignancy of the hematopoietic system that encompasses over 90 recognized subtypes, traditionally divided into non-Hodgkin lymphoma (NHL) and Hodgkin lymphoma (HL). According to GLOBOCAN 2022 data, approximately 0.08 million new cases of HL and 0.5 million new cases of NHL were diagnosed worldwide, with lymphoma accounting for 0.27 million deaths (2.8% of the 9.73 million total cancer-related deaths globally). NHL represents the more common form, with diffuse large B-cell lymphoma (DLBCL) being its most prevalent subtype, followed by follicular lymphoma (FL), mantle cell lymphoma (MCL), and marginal zone lymphoma (MZL).
The diagnostic challenge: Lymphoma typically presents as painless lymph node enlargement, but can also manifest without obvious symptoms, which contributes to misdiagnosis. The high biological heterogeneity among subtypes, combined with wide variation in five-year survival rates, makes accurate subclassification clinically essential. Treatment selection depends on subtype: indolent FL may be managed with watchful waiting, whereas DLBCL requires intensive chemoimmunotherapy. A single biopsy from one lesion site frequently fails to capture the full spatial and molecular heterogeneity of a disease that can involve the liver, lungs, soft tissues, and bone marrow simultaneously.
The role of PET/CT: 18F-fluorodeoxyglucose (18F-FDG) PET/CT has become indispensable in lymphoma management, offering quantitative data for disease burden assessment, treatment response evaluation, and prognostic stratification. Unlike conventional imaging, it captures metabolic activity across all lesion sites simultaneously. Radiomics, the computer-aided extraction of high-dimensional quantitative features from medical images, extends this capability by characterizing tumor heterogeneity through statistical descriptors that the human eye cannot perceive.
This 2025 review from Duan et al, published in Cancer Management and Research by researchers at Jining No.1 People's Hospital, Shandong First Medical University, synthesizes how AI-assisted PET/CT radiomics is being applied across the full lymphoma management pipeline: from differential diagnosis and subtype classification, through staging and risk stratification, to prognosis prediction and CAR-T therapy monitoring.
The fundamental radiomics workflow follows a six-stage pipeline: image acquisition and standardization, region labeling and tumor segmentation, feature extraction, feature selection and dimensionality reduction, model construction, and model evaluation. Each stage introduces potential sources of variability, which is why standardization frameworks such as the Image Biomarker Standardization Initiative (IBSI) and the Radiomics Quality Score (RQS) have been developed to ensure reproducibility across institutions and imaging platforms.
Feature categories: Radiomic features extracted from PET/CT images fall into several distinct families. First-order (histogram) features, including SUVmax, SUVmean, and SUVpeak, characterize gray-level distributions without spatial context. Second-order texture features, derived from matrices such as the Gray Level Co-occurrence Matrix (GLCM), Gray Level Run Length Matrix (GLRLM), Gray Level Size Zone Matrix (GLSZM), Gray Level Distance Zone Matrix (GLDZM), and Neighborhood Gray Tone Difference Matrix (NGTDM), quantify spatial organization and intensity hierarchies that reflect intratumoral heterogeneity. Model-based and transform-based features (using Fourier, Gabor, and Haar wavelet transformations) add further dimensionality by analyzing gray-level patterns in transformed mathematical spaces.
Machine learning and deep learning approaches: Machine learning (ML) algorithms, including support vector machines (SVM), random forests (RF), and decision trees (DT), are used to identify relationships between high-dimensional radiomic features and clinical outcomes. Models are evaluated using receiver operating characteristic (ROC) curves and area under the curve (AUC) metrics. Deep learning (DL), a subset of ML, uses multi-layer convolutional neural networks (CNNs) to learn representations directly from image data. The workflow progresses from classical feature engineering through supervised ML to end-to-end deep learning, with each approach offering different tradeoffs between interpretability, data requirements, and predictive power.
AI capabilities in imaging: AI's contributions to PET/CT analysis extend beyond feature extraction into automated tumor segmentation, multi-modal image registration, attenuation correction, artifact reduction, and computer-aided diagnosis. These capabilities improve both the quality of the imaging data fed into radiomics pipelines and the reliability of downstream predictions.
One of the first clinical applications of PET/CT radiomics in lymphoma is improving differential diagnosis, particularly in anatomical locations where lymphoma mimics other malignancies. Kong et al examined 77 patients (24 with central nervous system lymphoma and 53 with glioblastoma) and identified 13 PET radiomic features capable of distinguishing the two entities, with lymphoma showing higher SUV values across most intensity segments and greater quantitative heterogeneity. A study by Ou et al used 18F-FDG PET/CT radiomics combined with machine learning in 44 patients to separate breast lymphoma from breast cancer; the PETa model (combining clinical data, SUV measurements, and radiomic features) achieved AUC of 0.867 in training and 0.806 in validation, while the CT-based CTa model achieved AUC of 0.891 and 0.759, respectively.
Large-scale CNN classification: Sibille et al enrolled 629 patients (327 with lymphoma and 302 with lung cancer) in a deep CNN study using 18F-FDG PET/CT to discriminate between the two diagnoses. 18F-FDG PET alone achieved AUC of 0.97, and combining PET with CT further improved this to AUC of 0.98. Additional studies demonstrated PET/CT radiomics utility in differentiating primary central nervous system lymphoma from brain metastases (Cui et al, 2023: AUC 0.844-0.909 for density features), differentiating renal lymphoma from renal cell carcinoma (AUC 0.725-1.0), and distinguishing pancreatic lymphoma from pancreatic carcinoma (AUC 0.93).
Lymphoma subtype classification: Within the lymphoma category itself, radiomics enables non-invasive subtype differentiation that would otherwise require repeated invasive biopsies. Lovinfosse et al integrated age, sex, weight, and radiomic features from both original images and tumor-liver radiomics (TLR) in 420 patients (169 with sarcoidosis, 140 with HL, and 111 with DLBCL), using seven feature selection methods with four ML classifiers. The TLR lesion-based approach achieved AUC of 0.95 for HL vs. DLBCL discrimination, while the patient-based approach incorporating original radiomics and age yielded AUC of 0.86. A separate study using multiple-instance learning combined with SVM and random forest classifiers achieved 97.0% sensitivity and 94.1% positive predictive value for HL diagnosis evaluated at both volume-of-interest and patient levels.
Histological transformation detection: Richter transformation (RT), the conversion of chronic lymphocytic leukemia/small lymphocytic lymphoma (CLL/SLL) into aggressive B-cell lymphoma, occurs in roughly 2-10% of cases and carries a poor prognosis. A meta-analysis of 1,593 CLL patients established an optimal SUVmax threshold of 5 for detecting RT, demonstrating 86.8% sensitivity and 90.5% negative predictive value. Beyond SUVmax, metrics including lesion-to-liver SUV ratio (L-L SUVR) and lesion-to-blood-pool SUV ratio (L-BP SUVR) showed significant differences between RT and non-RT patients.
Accurate staging is foundational to treatment planning in both HL and NHL. NCCN and Society of Nuclear Medicine guidelines designate FDG-PET/CT as indispensable for lymphoma staging. Research shows it alters staging in 18-45% of FL patients, 3-45% of HL patients, and approximately 5% of DLBCL patients compared to conventional staging methods. The key challenge at the staging step is reliable detection of bone marrow involvement (BMI), which occurs in approximately 50% of NHL patients and up to 15% of HL patients. Bone marrow biopsy (BMB) remains the gold standard but is highly invasive, non-repeatable, and susceptible to false-negative results from random sampling errors.
Quantitative PET metrics for bone marrow assessment: Simple visual PET assessment of bone marrow has performed poorly, with sensitivity ranging from only 12-52% in mantle cell lymphoma cases. Quantitative threshold approaches have sought to identify cutoff values for cSUVmean (3D partial volume corrected mean), SUVmax, and SUVpeak that separate involved from uninvolved marrow. One study determined optimal cutoffs of 1.3, 2.1, and 1.7 for these three metrics respectively, achieving sensitivity/specificity pairings of 75.0%/85.7%, 87.5%/85.7%, and 87.5%/85.7%. However, a head-to-head comparison by Adams et al in 40 newly diagnosed DLBCL patients showed substantial overlap in these values between BMB-positive and BMB-negative groups, highlighting the ongoing controversy.
Radiomic features for BMI prediction: Aide et al extracted histogram, co-occurrence matrix, and size region matrix features from PET scans of 82 DLBCL patients. Among histogram features, SkewnessH emerged as the most accurate BMI predictor with 81.8% sensitivity and 81.7% specificity. Additional analyses identified code similarity and long-run emphasis as significant radiomic predictors of BMI. Mayerhoefer et al applied a multilayer perceptron neural network to SUV parameters combined with 16 co-occurring matrix texture features from 97 MCL patients. SUVs alone achieved AUC 0.66, radiomic features alone reached AUC 0.73, and integrating radiomic features with laboratory data pushed AUC to 0.81.
Risk stratification for treatment intensification: For DLBCL, Eertink et al extracted 490 radiomics features from baseline PET/CT of 317 patients using semi-automated segmentation (SUV threshold 4.0). Fusing radiomics and clinical features, particularly the combination of tumor-centric metrics (MTV, SUVpeak, Dmaxbulk) with patient-centric parameters (WHO performance status and age over 60 years), achieved AUC of 0.79 for identifying high-risk patients. Adding radiomics features to clinical features produced a 15% increase in positive predictive value. For HL, four risk categories defined by TMTV (threshold 147) and interim PET Deauville score showed 5-year PFS rates of 95%, 81.6%, 50%, and 25%, respectively.
Traditional prognostic systems for lymphoma, including the International Prognostic Index (IPI), revised IPI, and NCCN-IPI, rely on a small set of clinical variables, age, Ann Arbor stage, extranodal involvement, serum LDH, and performance status, that serve as indirect proxies for tumor burden and fail to capture metabolic or spatial heterogeneity. PET/CT-derived quantitative biomarkers offer direct measures of tumor biology that complement and in many settings outperform these clinical indices.
SUVmax limitations: SUVmax, reflecting the most metabolically active tumor region, is the most widely studied PET biomarker. In one study, DLBCL patients with SUVmax below 15 had 3-year OS of 90% versus 72% for those with SUVmax 15 or higher, with corresponding 3-year PFS of 90% versus 39%. However, large-scale evidence has challenged its reliability. A study of 258 stage I-II HL patients found SUVmax had negligible predictive value for PFS and OS. Variability in scanner resolution, acquisition protocols, PET reconstruction parameters, and cohort composition are all cited as reasons for inconsistent results.
MTV and TLG as tumor burden metrics: Metabolic tumor volume (MTV) and total lesion glycolysis (TLG) provide volumetric assessment of overall tumor burden. Song et al analyzed 169 stage II-III DLBCL patients and identified an optimal MTV cutoff of 220 cm3 using a SUV threshold of 2.5, with patients below this threshold showing significantly prolonged PFS and OS. In a separate DLBCL cohort, high TLG50 values above 415.5 were linked to 2-year PFS of 73% versus 92% and 2-year OS of 81% versus 93% in the low TLG50 group, with hazard ratios of 4.4 for PFS and 3.1 for OS. In the prospective Phase III GOYA study (1,418 patients), elevated TMTV (366 cm3) and TLG (3,004 g) were independently associated with inferior PFS in DLBCL.
Maximum tumor dissemination (Dmax): Dmax, defined as the maximum distance between the two farthest lesions, captures the spatial spread of multifocal disease. Cottereau et al validated Dmax as an independent predictor of PFS (HR=4.3) and OS (HR=3.7) in advanced HL. Gallamini et al found the most precise threshold for Dmax in predicting HL treatment outcomes to be 16.2 cm, with AUC of 0.62. Dmax is computationally simpler than MTV, requiring minimal contour dependency, and shows high reproducibility across imaging platforms.
Metabolic heterogeneity (MH): MH, quantified using the area under the cumulative SUV-volume histogram (AUC-CSH), reflects the distribution of FDG uptake within the dominant tumor lesion rather than just its peak or volume. In a prospective study of 103 primary mediastinal B-cell lymphoma patients, elevated MH independently predicted reduced PFS (HR=12.8) in stepwise Cox regression. Among DLBCL patients receiving R-CHOP, elevated MH in the high-MTV subgroup predicted significantly worse outcomes. For follicular lymphoma, a meta-analysis of 27 studies demonstrated MTV's prognostic value with a pooled hazard ratio of 3.05 for PFS.
Beyond volumetric biomarkers, structural and texture analysis of PET/CT images captures subvisual spatial patterns that reflect tumor biology at a finer scale. Second-order texture parameters derived from GLCM matrices and higher-order features quantify the spatial organization of FDG uptake intensity across the tumor volume, capturing characteristics such as homogeneity, contrast, correlation, and dissimilarity that correlate with tumor aggressiveness and treatment resistance.
Texture features for DLBCL event-free survival: Aide et al demonstrated that integrating MTV with second-order texture parameters (homogeneity, contrast, correlation, dissimilarity) and higher-order features (LZE, LZLGE, LZHGE, GLNU, ZP) significantly predicts 2-year event-free survival (2y-EFS) in DLBCL. Multivariate analysis identified LZHGE (long-zone high gray-level emphasis, a higher-order feature capturing large high-intensity regions) as an independent prognostic marker with HR=7.47. Ritter et al identified five key predictors for 2y-EFS in DLBCL: maximum diameter, NGTDM busyness, TLG, TMTV, and NGTDM coarseness. These NGTDM features capture the rate of intensity change between adjacent voxels, reflecting local texture complexity.
Machine learning models for DLBCL: Frood et al analyzed baseline PET/CT scans from 229 DLBCL patients treated with R-CHOP, training and optimizing logistic regression and six ML classifiers using 4-fold cross-validation. A ridge regression model combining clinical and radiomic features achieved mean training AUC of 0.77 plus or minus 0.02 and test AUC of 0.73, outperforming the MTV-only model (AUC 0.67). Capobianco et al used automated whole-body CNN segmentation in 301 DLBCL patients to calculate TMTVPARS, which showed strong agreement with expert semi-automated TMTVREF while maintaining equivalent predictive value for PFS and OS, suggesting that automated methods can substitute for time-consuming expert contouring.
Radiomic models for HL: Milgrom et al evaluated PET/CT radiomic parameters in 251 stage I-II HL patients at a tertiary cancer center. A machine learning model using the five most predictive features (SUVmax, volume, GLCM information measure Corr 1, information measure Corr 2, and GLCM mean variance) achieved AUC of 95.2% for predicting refractory disease. Frood et al analyzed 289 classical HL patients and found that a ridge regression model incorporating age and four radiomic features (PET flatness, PET major axis length, GLSZM, GLCM, and PET lbp-3D-m2) with 1.5x mean liver SUV segmentation achieved validation AUC of 0.79 plus or minus 0.01.
Chimeric antigen receptor T cell (CAR-T) therapy has emerged as a transformative approach for relapsed or refractory NHL, including DLBCL and classical HL. In this high-intensity treatment setting, PET/CT parameters play an increasingly defined role in predicting who will benefit, who will experience severe toxicity, and how durable the response will be. The modality offers a non-invasive window into pre-treatment disease burden and post-treatment metabolic activity across all disease sites simultaneously.
Tumor burden before CAR-T infusion: Baseline PET/CT metrics have demonstrated predictive value for outcomes following CAR-T treatment. Baseline SUVmax can predict progression-free survival following CAR-T infusion, and volumetric parameters, MTV and TLG, serve as potential prognostic indicators affecting both PFS and OS. A study examining CD30-targeted CAR-T cells for relapsed/refractory classical HL found that high MTV (60 mL) before lymphocyte depletion was associated with markedly inferior 1-year PFS rates of 14% in the high-MTV group versus 58% in the low-MTV group. In relapsed/refractory cHL patients treated with autologous stem cell transplantation, high MTV at relapse and before transplantation also predicted more severe PFS outcomes.
Toxicity prediction: The two most prevalent adverse effects of CAR-T cell therapy are cytokine release syndrome (CRS), characterized by systemic inflammatory response from rapid cytokine release by activated CAR-T cells, and neurotoxicity, manifesting as seizures and disorientation. Multiple PET/CT parameters, including SUVmax, SUVavg, MTV, and TLG, may help predict and assess the severity of these toxicities in CAR-T recipients. This represents a potentially actionable application: clinicians could use pre-treatment PET metrics to stratify toxicity risk and adjust monitoring protocols or prophylactic interventions accordingly.
Clinical implications: The integration of FDG PET/CT parameters into CAR-T therapy management offers a comprehensive strategy that extends beyond response assessment to encompass proactive toxicity management. As CAR-T therapy becomes more widely used for lymphoma, standardizing which PET/CT metrics are collected pre-treatment and how they inform patient selection or risk stratification protocols will be critical for extracting the full predictive value of this modality.
Image quality and standardization: The primary technical challenge in PET/CT radiomics is image quality control. Radiomic features are sensitive to variations in acquisition parameters, including scan duration, iteration and subset numbers, reconstruction type and algorithm, and spatial resolution. One study found that spatial resolution variations had the most pronounced impact across all sources of heterogeneity, with a coefficient of variation reaching 3.63. To address this, multiple standardization frameworks have been progressively implemented: the Radiomics Quality Score (RQS), the Image Biomarker Standardization Initiative (IBSI), the AI in Medical Imaging Checklist (CLAIM), and the Quantitative Imaging Biomarkers Alliance (QIBA). Achieving consensus on image acquisition protocols and analytical workflows remains an active area of effort across the field.
Retrospective designs and selection bias: The majority of lymphoma radiomics studies adopt retrospective single-center designs. Given lymphoma's subtype heterogeneity and the potential for histological transformation, performing pathological confirmation of each lesion under investigation is rarely feasible in retrospective cohorts. This introduces selection bias in region-of-interest (ROI) delineation and case inclusion, limiting the generalizability of models trained on data from a single institution's scanner and patient population. Large-scale prospective multicenter studies with independent external validation through multi-institutional collaborations are needed to establish clinical-grade model robustness.
Interpretability and deep learning limitations: CNN-based end-to-end models learn from raw imaging data without explicit feature engineering, but their decision logic is not transparently interpretable by clinicians. This limits trust and clinical adoption. While AI has demonstrated capabilities in refined attenuation correction, artifact-free reconstruction, and automated annotation, the black-box nature of deep learning models remains a barrier to regulatory approval and clinical translation. Developing explainability methods for radiomic and deep learning models, particularly in the lymphoma context, is a recognized priority.
Genomic integration and future opportunities: Clinicians are increasingly aware that single-site biopsy samples inadequately reflect patient-level genomic profiles across heterogeneous multifocal disease. Radiomics offers the potential to combine imaging phenotypes with genomic data including circulating tumor DNA (ctDNA) and MYC rearrangements to provide more complete biological characterization of each patient. Research has revealed correlations between lymphoma imaging phenotypes and gene expression patterns, creating opportunities for radiogenomics approaches that link PET/CT-derived imaging signatures to molecular subtypes and treatment sensitivity profiles. The authors conclude that integrating PET/CT imaging, radiomics, and genomic data represents the most promising direction for precision oncology in lymphoma management.